[GPU] Fix synchronization in the PagedAttention kernel when KV-cache rotation is enabled but skipped for the current iteration

sshlyapn · sshlyapn · commit 46fb2f514611 · 2025-01-23T18:02:42.000+04:00
diff --git a/src/plugins/intel_gpu/src/graph/impls/ocl/paged_attention.cpp b/src/plugins/intel_gpu/src/graph/impls/ocl/paged_attention.cpp
@@ -346,7 +346,7 @@ struct paged_attention_impl : multi_stage_primitive<paged_attention> {
 
         std::vector<event::ptr> res_events;
         std::vector<event::ptr> dep_events = events;
-        if (has_rotated_blocks) {
+        if (has_rotated_blocks && !_kernels_data[Stage::KV_CACHE_ROTATE].kernels[0].skip_execution) {
             execute_stage(dep_events, instance, res_events, Stage::KV_CACHE_ROTATE, is_mixed_mode);
             dep_events = res_events;
         }

Original file line number	Diff line number	Diff line change
`@@ -346,7 +346,7 @@ struct paged_attention_impl : multi_stage_primitive<paged_attention> {`
`346`	`346`
`347`	`347`	`std::vector<event::ptr> res_events;`
`348`	`348`	`std::vector<event::ptr> dep_events = events;`
`349`		`- if (has_rotated_blocks) {`
	`349`	`+ if (has_rotated_blocks && !_kernels_data[Stage::KV_CACHE_ROTATE].kernels[0].skip_execution) {`
`350`	`350`	`execute_stage(dep_events, instance, res_events, Stage::KV_CACHE_ROTATE, is_mixed_mode);`
`351`	`351`	`dep_events = res_events;`
`352`	`352`	`}`